AI systems are only as good as the training that goes into them – we pointed out a few examples of poorly trained systems last summer. Effective training can be divided into two parts: data and models.
The data is many things. There are utterances – all the words that an enterprise’s customers hand to them; what they said and how they responded. There are customer clickstreams – what did the customer do during and after that session they had, giving you hints and clues to their satisfaction. There is customer context – external data about that customer; plan or services they subscribe to, did they pay their bill, are they out of the country? whatever you’ve got, all that data you want to gather in and use that to understand your customers need the customer’s history, what are they doing in the past next best offer?
This is all data that is used by analysts, data scientists, content experts, and quality assurance departments to perform AI training and shape the care and service an enterprise can offer to its customers.
The other side is the training models, which extends beyond natural language processing. What’s going to help you refine your understanding of your customer’s needs? Training models need to determine customer satisfaction while simultaneously matching customer context. There is too much data for individual analysis – tens to hundreds of thousands of transactions every day. Enterprises need a system – a model – that automatically determines what a happy customer looks like through the data in these interactions. This system must be trained by linguists, machine learning experts, research scientists and architects.
In the above video, Wysdom CEO Ian Collins walks us through the differences between the two, how the combine to make up the full spectrum of AI Training, and provides a few real life examples.